The content is supposed to be a Parquet file containing the metadata of all parquet files
  in the Parquet Dataset, including metadata of each column chunk in each file in the hierarchy.
  See ParquetDataset in pyarrow: 
    https://arrow.apache.org/docs/python/generated/pyarrow.parquet.ParquetDataset.html.
  
  In practise, the "_metadata" parquet file contains all "RowGroupMetaData". 
  Each "RowGroupMetaData" contains a list of "ColumnChunkMetaData" and each "ColumnChunkMetaData" 
  contains a "file_path" attribute pointing to a file in the hierarchy 
  (together with the byte offset to quicky locate the chunk inside the file).
  
  Generating the file on-the-fly and not ahead-of-time, we are not able to produce quickly such metadata.
  The files "_metadata" and "_common_metadat" are generated by "pyarrow" and are not part of the
  parquet standard.
  See section "writting _metadata and _common_metadata" in:
    https://arrow.apache.org/docs/python/parquet.html
  "Note this is not a Parquet standard, but a convention set in practice by those frameworks."
  
  I am realy relunctant to the idea of making an IVOA standard based on practices of a few frameworks,
  especially since it would prevent from generating files on-the-fly.
  
  Also, in the case of Gaia DR3 with files of maximum 1_000_000 rows, the "_metadata" file is 94 MB 
  large. Does it mean that any query on the Gaia DR3 dataset would start by donwloading ~100MB
  of (meta)data?